Project description

The area surrounding the Frank Lake flux tower is composed of several vegetation types. These vegetation classes have been observed as a Hordeum, Bulrush, Barren, and an central island that is a mix of Hordeum and ??Puccinellia??. Data from the flux tower shows the prevailing wind directions are from the southwest and northeast. Therefore, understanding the vegetation composition of these sections will improve calculations of gas exchange measured by the flux tower.

Packages

Details of R software and package versions can be found in the renv.lock file of the project directory. The renv package provides a method for loading all packages need to recreate the work flow of this project.

Data

On October 14, 2022 Stew Rood used the DJI drone to fly over the Frank Lake study site. The resulting image was a RGB raster of approximately 3 cm spatial resolution and was compiled by the DroneDeploy software.

Due to the fine spatial resolution the initial file was cropped and downscaled to reduce the file size. A 200 m radius circle from the flux tower (center of the island) was selected for cropping to ensure enough data for future image classification. The raster was downscaled to the 30 cm using the aggregate function from the R package terra (v 1.6.17). 30 cm was selected as a single order of magnitude decrease in resolution from the initial image and therefore would not split mixes in the process of averaging.

The resulting image can be seen below. Note that the boardwalk, ATV path, and instruments have all been removed from the image by manually digitizing these elements in QGIS and converting their values to NA.

Processing

Creating radial sections

A series of sections originating from the flux tower were created to use as extraction polygons for the vegetation classes. These sections were determined by the angle and distance from the flux tower. Distances are set at 25 m increments up to 150 m while angles will be set at 22.5° intervals. Note that 0° is at the 12 o’clock position of the image and go clockwise. The buffer_wedge function from buffeRs (v 0.31) was used.

RGB indices

Following cleaning the initial RGB image, a series of indices were calculated using the rgb_indices function from the uavRst package (v 0.6.0). More details on the specific indices calculated can be found in the function help documentation. Here, the following indices were input into our image classification models: VVI, VARI, NDTI, RI, SCI, BI, SI, HI, TGI, GLI, NGRDI, GLAI, CI, SAT, SHP.

Plots of these indices can be seen below.

Image classification

To classify the vegetation present across the study area two methods were used.

  1. Manual delineation of RGB image
  2. CART model using RGB and calculated indices

The manual delineation approach used QGIS (v 3.24.2) to trace the outline of the island, Hordeum, and barren, and path areas. “Snap geometries to layer” function in QGIS was used to with 1 meter tolerance fill small gaps between adjacent classifications. The remaining unclassified space was then classified as Bulrush to limit added error of manually digitizing the larger area of the most dominant class.

A classification and regression tree (CART) model was used to classify pixels into each vegetation class using a subset of manually defined training points. Within each subset up to 50,000 training points were selected. The points were input into the rpart function from the rpart package (v 4.1.16) using method “class” and minsplit of 5.

Since the island is a combination of two plant species the model struggles to accurately predict the vegetation present. Therefore, two CART models were used, 1) where the island is included in the training points as it’s own class and 2) where the island is set to NA and removed from the model. In the latter option the island was filled in using the manually delineated polygon since this feature is easily defined by visual observation.

The CART models were evaluated using k-fold validations. The data data was split into k=5 groups and the model is refit with one group being used for model testing while the remaining will be used for model training. Following this evaluation, the island included CART model had an overall accuracy of 0.84 while the island excluded CART model had an overall accuracy of 0.95.

Each method can be seen below in addition to the RGB image and the section overlay

Extraction

The vegetation composition of each section was then extracted for each of the three methods used (manual, CART, and CART + manual). Since the sections do not align perfectly with the pixels the proportion of each pixel within each section was calculated and multiplied by the pixel resolution (0.2375003 m) to get the area. Each vegetation type (Hordeum, Bulrush, Barren, and Island) were then summed to get the total area of that class in each section.

As noted the boardwalk was excluded from these areas as it would not contribute to the vegetation cover. Total area of each section and proportional composition were also calculated. Sections are denoted by the angle and distance from the flux tower. For example, angle_112.5_dist_100 indicates the section 112.5° from 12 o’clock and 100 m from the flux tower.